InfoGAIL: Interpretable Imitation Learning from Visual Demonstrations

By Yunzhu Li, Jiaming Song, Stefano Ermon

Introduction

Modified codebase of TORCS, with the ability to extract dashboard views.

InfoGAIL implementation, attached with two examples: pass & turn.

Citing InfoGAIL

If you find this codebase useful in your research, please consider citing:

@article{li2017inferring,
    title={InfoGAIL: Interpretable Imitation Learning from Visual Demonstrations},
    author={Li, Yunzhu and Song, Jiaming and Ermon, Stefano},
    journal={arXiv preprint arXiv:1703.08840},
    year={2017}
}

Requirements

Python 2.7
Tensorflow 0.12.1
Keras 1.2.2
xautomation

Install and Configure TORCS

In the following section, we only show the commands for running example 0 (indicating pass). Users can replace 0 with 1 to run example 1 (indicating turn).

Install all the dependencies of TORCS, including but not limited to:

sudo apt-get install libplib-dev
sudo apt-get install libopenal-dev
sudo apt-get install libalut-dev

See http://torcs.sourceforge.net/ for more information

cd to the torcs-1.3.4/ folder, type the following commands to install TORCS:

./configure
make
sudo make install
sudo make datainstall

Default installation directories are:

/usr/local/bin
/usr/local/lib/torcs
/usr/local/share/games/torcs

Run the torcs command to play TORCS.

Copy the modified tracks files in the modified_tracks/ folder to the torcs folder /usr/local/share/games/torcs/tracks/road
Type the following commands to configure the running environment:

rm -rf ~/.torcs
cp -r torcs_config_ai_0 ~/.torcs

Download training data and pretrained weights

Training data

example 0 (pass) DropBox
example 1 (turn) DropBox

Pretrained weights

example 0 DropBox
example 1 DropBox

Run with pretrained weights

cd to wgail_info_0/
open drive.py and edit line 20-21 in concert with downloaded data and weights
change variable code in line 14 into different values (0 or 1) to observe different behaviors
type the python drive.py to run pass with pretrained weights

Run pass with different latent codes (0 or 1):

Run turn with different latent codes (0 or 1):

Training

cd to wgail_info_0/
open wgail_info.py and edit line 17-19 in concert with downloaded data and weights
open models.py and edit line 508 to specify the place to store the weights, edit line 422 to specify the place to store the log file
type python wgail_info.py to train on pass, note that the weights trained via behavior cloning are used to initialize the policy network

Misc.

Track selection:
- pass: chenyi-Street 1
- turn: chenyi-Brondehach
torcs-1.3.4 in this repository is a modified version of the original codebase released by Bernhard Wymann. This version holds the ability of extracting and transmitting visual information.